Which is better, GPT-5.6 Sol or Kimi K2.5?

Across 50 shared benchmarks tracked on Vector Wire, GPT-5.6 Sol leads on 45 and Kimi K2.5 leads on 5 — GPT-5.6 Sol leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

GPT-5.6 Sol vs Kimi K2.5

OpenAIvsMoonshot50 shared benchmarks45–5 head-to-head

Benchmark	GPT-5.6 Sol	Kimi K2.5
AA Agentic Index	57.8	52.8
AA Intelligence	61	36
AA-LCR	77.7	73
AA-Omniscience	22	-7.3
aime_2026	99.9	95.8
arc_agi_1	97.5	65.3
ARC-AGI-2	92.5	11.8
arena_elo	1482	1450
arena_text_factuality	1474	1445
Artificial Analysis Coding Index	78.3	46.8
baby_vision_with_python	88.9	40.5
browsecomp	90.4	74.9
BrowseComp (w/ Ctx)	89.4	74.9
charxiv_rq	87.8	78.7
coding_arena_elo	1619	1436
critpt	32.3	3.1
cybergym	83.6	41.3
FORTRESS (Adversarial)	82.4	54.1
FORTRESS (Benign)	98.1	98.3
gdpval	61.1	38.3
GDPval-AA v2	1748	1009
Global-MMLU-Lite	91.8	84
GPQA Diamond	94.6	87.9
HLE	49.5	50.2
HLE (with tools)	58	51.8
IFBench	72.7	70.2
longbench_v2	67.1	61
mathvision	95.8	84.2
MCP Atlas	83.6	64
MMMU-Pro	84.6	78.5
MMVU	81.2	80.4
OmniScience Accuracy	59.4	35.2
OmniScience Non-Hallucination	10.6	50
OSWorld-Verified	83	63.3
PaperBench	90.5	63.5
ResearchRubrics	73.8	59.5
scicode	56.9	49
simplebench	64.8	46.8
simpleqa_verified	71.6	36.9
strongreject	98.5	99.5
SWE-bench Pro	64.6	53.8
SWEBench Pro (Public)	64.6	50.7
Tau 3 Banking	33	14.2
TauBench V3 - Banking	44.3	14.2
Terminal Bench 2.1 (Best Harness)	89.5	51.3
Terminal-Bench 2.1	89.5	45.7
Terminal-Bench Hard	65.9	34.8
toolathlon	58	27.8
Video-MME	89.5	87.4
τ²-Bench Telecom (AA run)	85.1	95.9

Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.